Обсуждение: FreeBSD and large shared_buffers a no-go?
I'm running PostgreSQL 8.3.5 on a FreeBSD 7.1/amd64 system with 8GB of RAM and two quad-core Xeon CPUs. The data filesystem is on a battery-backed RAID-10 system. This is a dedicated server so I want to commit all resources to PostgreSQL alone. The database will get hit with a lot of small, quick queries with a few complex 10-second ones thrown in, and gets hourly bulk uploads to the tune of a few million rows.
This is actually an upgrade from older hardware that handled the load just fine, so I know this system will perform well. My biggest concern is getting the best performance for my boss's money.
I'm using the default postgresql.conf with the following additions:
max_connections = 400
listen_addresses = '*'
shared_buffers = 2GB
temp_buffers = 32MB
work_mem = 64MB
maintenance_work_mem = 256MB
max_stack_depth = 500MB
max_fsm_pages = 204800
full_page_writes = off
wal_buffers = 1MB
commit_delay = 100000
checkpoint_segments = 32
random_page_cost = 2.0
effective_cache_size = 4GB
default_statistics_target = 100
log_connections = on
log_disconnections = on
log_min_duration_statement = 5000
log_statement = 'ddl'
Now, what's confusing me is that I've set shmmax to 3GB and shmall to 3GB/4096 (the page size):
$ sysctl kern.ipc.shmmax
kern.ipc.shmmax: 3221225472
$ sysctl kern.ipc.shmall
kern.ipc.shmall: 786432
$ sysctl hw.pagesize
hw.pagesize: 4096
However, when shared_buffers is 2GB (one fourth of 8GB of RAM), PostgreSQL's startup fails with a call to allocated shared memory:
Jan 7 11:39:24 db1 postgres[60872]: [1-1] FATAL: could not create shared memory segment: Cannot allocate memory
Jan 7 11:39:24 db1 postgres[60872]: [1-2] DETAIL: Failed system call was shmget(key=5432001, size=2209497088, 03600).
Jan 7 11:39:24 db1 postgres[60872]: [1-3] HINT: This error usually means that PostgreSQL's request for a shared memory segment exceeded available memory or swap space.
Jan 7 11:39:24 db1 postgres[60872]: [1-4] To reduce the request size (currently 2209497088 bytes), reduce PostgreSQL's shared_buffers parameter (currently 262144) and/or
Jan 7 11:39:24 db1 postgres[60872]: [1-5] its max_connections parameter (currently 403).
Jan 7 11:39:24 db1 postgres[60872]: [1-6] The PostgreSQL documentation contains more information about shared memory configuration.
First, shmget is asking for a lot less than shmmax - why is it failing?
Second, does the "one fourth of RAM" rule of thumb still apply on systems with decent amounts of memory?
Third, is there anything else I can be doing to take advantage of this RAM and 8 CPU cores?
Thanks!
--
Kirk Strauser
Kirk Strauser <kirk@strauser.com> writes: > First, shmget is asking for a lot less than shmmax - why is it failing? Check to see if things work as expected when you have shmmax and shmall set to a shade less than 2GB and fail when they are a shade more. If so, it would seem there's a signed-integer-overflow bug somewhere in the kernel's handling of shmem requests ... which would be a reportable kernel bug. regards, tom lane
On Jan 7, 2009, at 12:30 PM, Tom Lane wrote:
Kirk Strauser <kirk@strauser.com> writes:First, shmget is asking for a lot less than shmmax - why is it failing?
Check to see if things work as expected when you have shmmax and shmall
set to a shade less than 2GB and fail when they are a shade more. If
so, it would seem there's a signed-integer-overflow bug somewhere in the
kernel's handling of shmem requests ... which would be a reportable
kernel bug.
--
Kirk Strauser
On Jan 7, 2009, at 12:30 PM, Tom Lane wrote:
Kirk Strauser <kirk@strauser.com> writes:First, shmget is asking for a lot less than shmmax - why is it failing?
Check to see if things work as expected when you have shmmax and shmall
set to a shade less than 2GB and fail when they are a shade more. If
so, it would seem there's a signed-integer-overflow bug somewhere in the
kernel's handling of shmem requests ... which would be a reportable
kernel bug.
BTW, that's at http://www.freebsd.org/cgi/query-pr.cgi?pr=130274 for anyone who wants to follow along.
--
Kirk Strauser